AI safety tests have a new problem: Models are now faking their own reasoning traces the-decoder.com